Driving Cycle and Road Grade on-board prediction for the optimal energy management in EV-PHEVs Borja Heriz Revuelta borja.heriz@tecnalia.com Alternative Powertrain Researcher Tecnalia Transport Business Unit
Contents 1. Objective 2. Precedents 3. Proposal 4. The idea 5. Implementation 6. Results 7. Conclusions
Objective DRIVER FACTORS experience, age, physical conditions, etc. DRIVING CONDITIONS ROAD FACTORS urban, extra urban, highway, etc. TRAFFIC FACTORS congestion, accidents, peak/offpeak hours, etc. f (nonlinear) DRIVING CYCLE Vehicle speed (t) WEATHER FACTORS temperature, rain, snow, visibility, etc. Driving Cycle of Martorell-Barcelona trip. Courtesy of SEAT. VEHICLE FACTORS power/weight, geometry, age, etc. Driving cycle and road grade on-line prediction over a time horizon.
Precedents Research activities only focused on driving conditions (DCOs) real-time prediction. 1. Obtaining some statistical parameters of the recent past vehicle speed that feed a classification technique (Fuzzy Logic, ANN, ). 2. Enhancing them with digital maps information and historical route data for building a Reference Driving Cycle (RCD). No approach model the driver/vehicle behavior. Penalizing the prediction extra up to 15% fuel consumption reduction can be achieved.
Proposal Use an ANN for modeling the driver/vehicle behavior. Processing the deviation in the trip/distance domain between a RDC and the vehicle speed. It is only provoked by the driver behavior. The RDC is dynamically and easily constructed using only road/route information: Static GIS (Geographical Information System, digital maps) Dynamic GIS + V2X
The idea in 3 steps Real-Time (vehicle in movement) 1. RECORDING *Just once. At start 2. TRAINING Real-Time (vehicle in movement) 1. Recording RDCs and driving cycles during usual journeys. In the background. 2. Neural Network Training Based on the speed deviation between the RDC and the vehicle speed in the trip/distance domain(not time domain). Off-line task. 3. PREDICTION EXECUTION 1. RECORDING
3. Prediction Execution (I) The idea in 3 steps
3. Prediction Execution (II) The idea in 3 steps
Neural Network NARX (Non-linear Autoregressive with exogenous inputs). Widely used in nonlinear time series predictions. Multilayer net with two delayed inputs, a sigmoid-based hidden layer (7-15 neurons) and pure linear-based output layer (1 neuron). Training methods: Bayesian regularization, Levenberg- Marquardt. Low computational efforts once trained embedded into MCUs and DSPs.
Off-line work MATLAB-Based GUI for training and simulation. On-line work Rapid prototyping in a MicroAutobox from dspace. Tested in a real SEAT vehicle. CANBuscommunication for receiving vehicle speed (road/route information not available yet, thinking in GIS + ADASYS Protocol). Prototype Implementation
Results 1. Simulation results based on real speed data. Predictions for 4 and 12 kilometric points with 2km prediction horizon
Results 2. In-vehicle testing using a SEAT vehicle. Trip: Martorell-Sant Joan Despí Prediction horizon: 2km Prediction Mean Error lower than 10 km/h
Conclusions Driving cycles in the surroundings of Barcelona have been on-line predicted using advanced computational techniques with an average error less than 10km/h. Good performances obtained using prediction horizons ranged from 1 to 5km. Driving cycles predictions could be used for optimizing different vehicle functions. Among others, energy management in PHEVs, thermal management, transmission controls, eco-driving ADAS.
Acknowledgements This work has been realized in cooperation with the SEAT Technical Center under the scope of the CENIT VERDE research project granted by the Spanish Ministry of Economy and Competitiveness.